Image Analysis and Stereology · Published 2025-11-27 · DOI 10.5566/ias.3695
Accurate malignancy prediction in breast ultrasound imaging is challenged by limited annotated data, high inter-observer variability, and inherent noise in sonographic textures. To address these limitations, we propose a deep learning framework that synergistically integrates multiscale feature fusion and self-supervised learning (SSL) to improve diagnostic performance while minimizing reliance on labeled datasets. The architecture employs a hierarchical convolutional backbone with multiscale feature extractors that capture both coarse contextual semantics and fine-grained morphological cues of lesions. Features across multiple receptive fields are fused via a top-down multiscale fusion strategy using bilinear upsampling and channel concatenation, enhancing the model’s ability to localize and characterize malignant regions. We applied a self-supervised contrastive learning approach tailored for medical ultrasound, incorporating spatial transformation invariance and anatomical context preservation to learn domain-relevant representations from unlabeled data. The pretrained encoder is fine-tuned with a supervised classification head using a limited set of annotated images. Extensive experiments on two publicly available breast ultrasound datasets demonstrate that our model achieves higher performance over state-of-the-art baselines, yielding significant improvements in AUC, F1-score, and sensitivity. Ablation studies confirm the individual and combined efficacy of the multiscale fusion and SSL modules. This work establishes a scalable and label-efficient pipeline for ultrasound-based malignancy prediction, with implications for real-time clinical decision support.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →